diff mbox series

[v15,08/13] drm/i915/perf: implement active wait for noa configurations

Message ID 20190906093256.26198-9-lionel.g.landwerlin@intel.com (mailing list archive)
State New, archived
Headers show
Series drm/i915: Vulkan performance query support | expand

Commit Message

Lionel Landwerlin Sept. 6, 2019, 9:32 a.m. UTC
NOA configuration take some amount of time to apply. That amount of
time depends on the size of the GT. There is no documented time for
this. For example, past experimentations with powergating
configuration changes seem to indicate a 60~70us delay. We go with
500us as default for now which should be over the required amount of
time (according to HW architects).

v2: Don't forget to save/restore registers used for the wait (Chris)

v3: Name used CS_GPR registers (Chris)
    Fix compile issue due to rebase (Lionel)

v4: Fix save/restore helpers (Umesh)

v5: Move noa_wait from drm_i915_private to i915_perf_stream (Lionel)

v6: Add missing struct declarations in i915_perf.h

Signed-off-by: Lionel Landwerlin <lionel.g.landwerlin@intel.com>
Reviewed-by: Chris Wilson <chris@chris-wilson.co.uk> (v4)
---
 drivers/gpu/drm/i915/gt/intel_gpu_commands.h |  24 ++
 drivers/gpu/drm/i915/gt/intel_gt_types.h     |   5 +
 drivers/gpu/drm/i915/i915_debugfs.c          |  31 +++
 drivers/gpu/drm/i915/i915_drv.h              |   2 +
 drivers/gpu/drm/i915/i915_perf.c             | 234 ++++++++++++++++++-
 drivers/gpu/drm/i915/i915_perf_types.h       |   6 +
 drivers/gpu/drm/i915/i915_reg.h              |   4 +-
 7 files changed, 302 insertions(+), 4 deletions(-)
diff mbox series

Patch

diff --git a/drivers/gpu/drm/i915/gt/intel_gpu_commands.h b/drivers/gpu/drm/i915/gt/intel_gpu_commands.h
index a7f1377a54a2..f133f8dbacb1 100644
--- a/drivers/gpu/drm/i915/gt/intel_gpu_commands.h
+++ b/drivers/gpu/drm/i915/gt/intel_gpu_commands.h
@@ -158,6 +158,7 @@ 
 #define   MI_BATCH_GTT		    (2<<6) /* aliased with (1<<7) on gen4 */
 #define MI_BATCH_BUFFER_START_GEN8	MI_INSTR(0x31, 1)
 #define   MI_BATCH_RESOURCE_STREAMER (1<<10)
+#define   MI_BATCH_PREDICATE         (1 << 15) /* HSW+ on RCS only*/
 
 /*
  * 3D instructions used by the kernel
@@ -236,6 +237,29 @@ 
 #define   PIPE_CONTROL_DEPTH_CACHE_FLUSH		(1<<0)
 #define   PIPE_CONTROL_GLOBAL_GTT (1<<2) /* in addr dword */
 
+#define MI_MATH(x) MI_INSTR(0x1a, (x)-1)
+#define   MI_ALU_OP(op, src1, src2) (((op) << 20) | ((src1) << 10) | (src2))
+/* operands */
+#define   MI_ALU_OP_NOOP     0
+#define   MI_ALU_OP_LOAD     128
+#define   MI_ALU_OP_LOADINV  1152
+#define   MI_ALU_OP_LOAD0    129
+#define   MI_ALU_OP_LOAD1    1153
+#define   MI_ALU_OP_ADD      256
+#define   MI_ALU_OP_SUB      257
+#define   MI_ALU_OP_AND      258
+#define   MI_ALU_OP_OR       259
+#define   MI_ALU_OP_XOR      260
+#define   MI_ALU_OP_STORE    384
+#define   MI_ALU_OP_STOREINV 1408
+/* sources */
+#define   MI_ALU_SRC_REG(x)  (x) /* 0 -> 15 */
+#define   MI_ALU_SRC_SRCA    32
+#define   MI_ALU_SRC_SRCB    33
+#define   MI_ALU_SRC_ACCU    49
+#define   MI_ALU_SRC_ZF      50
+#define   MI_ALU_SRC_CF      51
+
 /*
  * Commands used only by the command parser
  */
diff --git a/drivers/gpu/drm/i915/gt/intel_gt_types.h b/drivers/gpu/drm/i915/gt/intel_gt_types.h
index dc295c196d11..f752b6cf9ea1 100644
--- a/drivers/gpu/drm/i915/gt/intel_gt_types.h
+++ b/drivers/gpu/drm/i915/gt/intel_gt_types.h
@@ -97,6 +97,11 @@  enum intel_gt_scratch_field {
 	/* 8 bytes */
 	INTEL_GT_SCRATCH_FIELD_COHERENTL3_WA = 256,
 
+	/* 6 * 8 bytes */
+	INTEL_GT_SCRATCH_FIELD_PERF_CS_GPR = 2048,
+
+	/* 4 bytes */
+	INTEL_GT_SCRATCH_FIELD_PERF_PREDICATE_RESULT_1 = 2096,
 };
 
 #endif /* __INTEL_GT_TYPES_H__ */
diff --git a/drivers/gpu/drm/i915/i915_debugfs.c b/drivers/gpu/drm/i915/i915_debugfs.c
index 9798f27a697a..8bb15f899d97 100644
--- a/drivers/gpu/drm/i915/i915_debugfs.c
+++ b/drivers/gpu/drm/i915/i915_debugfs.c
@@ -3572,6 +3572,36 @@  DEFINE_SIMPLE_ATTRIBUTE(i915_wedged_fops,
 			i915_wedged_get, i915_wedged_set,
 			"%llu\n");
 
+static int
+i915_perf_noa_delay_set(void *data, u64 val)
+{
+	struct drm_i915_private *i915 = data;
+
+	/* This would lead to infinite waits as we're doing timestamp
+	 * difference on the CS with only 32bits.
+	 */
+	if (val > mul_u32_u32(U32_MAX, RUNTIME_INFO(i915)->cs_timestamp_frequency_khz))
+		return -EINVAL;
+
+	atomic64_set(&i915->perf.noa_programming_delay, val);
+	return 0;
+}
+
+static int
+i915_perf_noa_delay_get(void *data, u64 *val)
+{
+	struct drm_i915_private *i915 = data;
+
+	*val = atomic64_read(&i915->perf.noa_programming_delay);
+	return 0;
+}
+
+DEFINE_SIMPLE_ATTRIBUTE(i915_perf_noa_delay_fops,
+			i915_perf_noa_delay_get,
+			i915_perf_noa_delay_set,
+			"%llu\n");
+
+
 #define DROP_UNBOUND	BIT(0)
 #define DROP_BOUND	BIT(1)
 #define DROP_RETIRE	BIT(2)
@@ -4348,6 +4378,7 @@  static const struct i915_debugfs_files {
 	const char *name;
 	const struct file_operations *fops;
 } i915_debugfs_files[] = {
+	{"i915_perf_noa_delay", &i915_perf_noa_delay_fops},
 	{"i915_wedged", &i915_wedged_fops},
 	{"i915_cache_sharing", &i915_cache_sharing_fops},
 	{"i915_gem_drop_caches", &i915_drop_caches_fops},
diff --git a/drivers/gpu/drm/i915/i915_drv.h b/drivers/gpu/drm/i915/i915_drv.h
index 7eb31923cde9..2c6f37219dff 100644
--- a/drivers/gpu/drm/i915/i915_drv.h
+++ b/drivers/gpu/drm/i915/i915_drv.h
@@ -1403,6 +1403,8 @@  struct drm_i915_private {
 
 		struct i915_oa_ops ops;
 		const struct i915_oa_format *oa_formats;
+
+		atomic64_t noa_programming_delay;
 	} perf;
 
 	/* Abstract the submission mechanism (legacy ringbuffer or execlists) away */
diff --git a/drivers/gpu/drm/i915/i915_perf.c b/drivers/gpu/drm/i915/i915_perf.c
index c9d0de3050fb..f2b778d84b52 100644
--- a/drivers/gpu/drm/i915/i915_perf.c
+++ b/drivers/gpu/drm/i915/i915_perf.c
@@ -197,6 +197,7 @@ 
 
 #include "gem/i915_gem_context.h"
 #include "gem/i915_gem_pm.h"
+#include "gt/intel_gt.h"
 #include "gt/intel_lrc_reg.h"
 
 #include "i915_drv.h"
@@ -408,6 +409,7 @@  static u32 *write_cs_mi_lri(u32 *cs, const struct i915_oa_reg *reg_data, u32 n_r
 }
 
 static struct i915_oa_config_bo* alloc_oa_config_buffer(struct drm_i915_private *i915,
+							struct i915_vma *noa_wait,
 							struct i915_oa_config *oa_config)
 {
 	struct i915_oa_config_bo *oa_bo;
@@ -436,7 +438,7 @@  static struct i915_oa_config_bo* alloc_oa_config_buffer(struct drm_i915_private
 					      MI_LOAD_REGISTER_IMM_MAX_REGS) * 4;
 		config_length += oa_config->flex_regs_len * 8;
 	}
-	config_length += 4; /* MI_BATCH_BUFFER_END */
+	config_length += 12; /* MI_BATCH_BUFFER_START into noa_wait loop */
 	config_length = ALIGN(config_length, I915_GTT_PAGE_SIZE);
 
 	oa_bo->bo = i915_gem_object_create_shmem(i915, config_length);
@@ -455,7 +457,12 @@  static struct i915_oa_config_bo* alloc_oa_config_buffer(struct drm_i915_private
 	cs = write_cs_mi_lri(cs, oa_config->b_counter_regs, oa_config->b_counter_regs_len);
 	cs = write_cs_mi_lri(cs, oa_config->flex_regs, oa_config->flex_regs_len);
 
-	*cs++ = MI_BATCH_BUFFER_END;
+
+	/* Jump into the NOA wait busy loop. */
+	*cs++ = (INTEL_GEN(i915) < 8 ?
+		 MI_BATCH_BUFFER_START : MI_BATCH_BUFFER_START_GEN8);
+	*cs++ = i915_ggtt_offset(noa_wait);
+	*cs++ = 0;
 
 	i915_gem_object_flush_map(oa_bo->bo);
 	i915_gem_object_unpin_map(oa_bo->bo);
@@ -550,7 +557,9 @@  int i915_perf_get_oa_config_and_bo(struct i915_perf_stream *stream,
 		mutex_unlock(&stream->config_mutex);
 
 		if (!oa_bo) {
-			oa_bo = alloc_oa_config_buffer(i915, oa_config);
+			oa_bo = alloc_oa_config_buffer(i915,
+						       stream->noa_wait,
+						       oa_config);
 			if (IS_ERR(oa_bo)) {
 				err = PTR_ERR(oa_bo);
 				goto err;
@@ -1524,6 +1533,16 @@  free_oa_buffer(struct i915_perf_stream *stream)
 	stream->oa_buffer.vaddr = NULL;
 }
 
+static void
+free_noa_wait(struct i915_perf_stream *stream)
+{
+	struct drm_i915_private *i915 = stream->dev_priv;
+
+	mutex_lock(&i915->drm.struct_mutex);
+	i915_vma_unpin_and_release(&stream->noa_wait, 0);
+	mutex_unlock(&i915->drm.struct_mutex);
+}
+
 static void
 free_oa_configs(struct i915_perf_stream *stream)
 {
@@ -1552,6 +1571,7 @@  static void i915_oa_stream_destroy(struct i915_perf_stream *stream)
 	mutex_unlock(&dev_priv->drm.struct_mutex);
 
 	free_oa_buffer(stream);
+	free_noa_wait(stream);
 
 	intel_uncore_forcewake_put(&dev_priv->uncore, FORCEWAKE_ALL);
 	intel_runtime_pm_put(&dev_priv->runtime_pm, stream->wakeref);
@@ -1731,6 +1751,202 @@  static int alloc_oa_buffer(struct i915_perf_stream *stream)
 	return ret;
 }
 
+static u32 *save_restore_register(struct drm_i915_private *i915, u32 *cs,
+				  bool save, i915_reg_t reg, u32 offset,
+				  u32 dword_count)
+{
+	uint32_t d;
+
+	for (d = 0; d < dword_count; d++) {
+		if (save) {
+			*cs++ = INTEL_GEN(i915) >= 8 ?
+				MI_STORE_REGISTER_MEM_GEN8 :
+				MI_STORE_REGISTER_MEM;
+		} else  {
+			*cs++ = INTEL_GEN(i915) >= 8 ?
+				MI_LOAD_REGISTER_MEM_GEN8 :
+				MI_LOAD_REGISTER_MEM;
+		}
+		*cs++ = i915_mmio_reg_offset(reg) + 4 * d;
+		*cs++ = intel_gt_scratch_offset(&i915->gt, offset) + 4 * d;
+		*cs++ = 0;
+	}
+
+	return cs;
+}
+
+static int alloc_noa_wait(struct i915_perf_stream *stream)
+{
+	struct drm_i915_private *i915 = stream->dev_priv;
+	struct drm_i915_gem_object *bo;
+	struct i915_vma *vma;
+	const u64 delay_ticks = 0xffffffffffffffff -
+		DIV64_U64_ROUND_UP(
+			atomic64_read(&i915->perf.noa_programming_delay) *
+			RUNTIME_INFO(i915)->cs_timestamp_frequency_khz,
+			1000000ull);
+	u32 *batch, *ts0, *cs, *jump;
+	int ret, i;
+	enum { START_TS, NOW_TS, DELTA_TS, JUMP_PREDICATE, DELTA_TARGET, N_CS_GPR };
+
+	bo = i915_gem_object_create_internal(i915, 4096);
+	if (IS_ERR(bo)) {
+		DRM_ERROR("Failed to allocate NOA wait batchbuffer\n");
+		return PTR_ERR(bo);
+	}
+
+	/*
+	 * We pin in GGTT because we jump into this buffer now because
+	 * multiple OA config BOs will have a jump to this address and it
+	 * needs to be fixed during the lifetime of the i915/perf stream.
+	 */
+	vma = i915_gem_object_ggtt_pin(bo, NULL, 0, 4096, 0);
+	if (IS_ERR(vma)) {
+		ret = PTR_ERR(vma);
+		goto err_unref;
+	}
+
+	batch = cs = i915_gem_object_pin_map(bo, I915_MAP_WB);
+	if (IS_ERR(batch)) {
+		ret = PTR_ERR(batch);
+		goto err_unpin;
+	}
+
+	/* Save registers. */
+	for (i = 0; i < N_CS_GPR; i++) {
+		cs = save_restore_register(
+			i915, cs, true /* save */, HSW_CS_GPR(i),
+			INTEL_GT_SCRATCH_FIELD_PERF_CS_GPR + 8 * i, 2);
+	}
+	cs = save_restore_register(
+		i915, cs, true /* save */, MI_PREDICATE_RESULT_1,
+		INTEL_GT_SCRATCH_FIELD_PERF_PREDICATE_RESULT_1, 1);
+
+	/* First timestamp snapshot location. */
+	ts0 = cs;
+
+	/*
+	 * Initial snapshot of the timestamp register to implement the wait.
+	 * We work with 32b values, so clear out the top 32b bits of the
+	 * register because the ALU works 64bits.
+	 */
+	*cs++ = MI_LOAD_REGISTER_IMM(1);
+	*cs++ = i915_mmio_reg_offset(HSW_CS_GPR(START_TS)) + 4;
+	*cs++ = 0;
+	*cs++ = MI_LOAD_REGISTER_REG | (3 - 2);
+	*cs++ = i915_mmio_reg_offset(RING_TIMESTAMP(RENDER_RING_BASE));
+	*cs++ = i915_mmio_reg_offset(HSW_CS_GPR(START_TS));
+
+	/*
+	 * This is the location we're going to jump back into until the
+	 * required amount of time has passed.
+	 */
+	jump = cs;
+
+	/*
+	 * Take another snapshot of the timestamp register. Take care to clear
+	 * up the top 32bits of CS_GPR(1) as we're using it for other
+	 * operations below.
+	 */
+	*cs++ = MI_LOAD_REGISTER_IMM(1);
+	*cs++ = i915_mmio_reg_offset(HSW_CS_GPR(NOW_TS)) + 4;
+	*cs++ = 0;
+	*cs++ = MI_LOAD_REGISTER_REG | (3 - 2);
+	*cs++ = i915_mmio_reg_offset(RING_TIMESTAMP(RENDER_RING_BASE));
+	*cs++ = i915_mmio_reg_offset(HSW_CS_GPR(NOW_TS));
+
+	/*
+	 * Do a diff between the 2 timestamps and store the result back into
+	 * CS_GPR(1).
+	 */
+	*cs++ = MI_MATH(5);
+	*cs++ = MI_ALU_OP(MI_ALU_OP_LOAD, MI_ALU_SRC_SRCA, MI_ALU_SRC_REG(NOW_TS));
+	*cs++ = MI_ALU_OP(MI_ALU_OP_LOAD, MI_ALU_SRC_SRCB, MI_ALU_SRC_REG(START_TS));
+	*cs++ = MI_ALU_OP(MI_ALU_OP_SUB, 0, 0);
+	*cs++ = MI_ALU_OP(MI_ALU_OP_STORE, MI_ALU_SRC_REG(DELTA_TS), MI_ALU_SRC_ACCU);
+	*cs++ = MI_ALU_OP(MI_ALU_OP_STORE, MI_ALU_SRC_REG(JUMP_PREDICATE), MI_ALU_SRC_CF);
+
+	/*
+	 * Transfer the carry flag (set to 1 if ts1 < ts0, meaning the
+	 * timestamp have rolled over the 32bits) into the predicate register
+	 * to be used for the predicated jump.
+	 */
+	*cs++ = MI_LOAD_REGISTER_REG | (3 - 2);
+	*cs++ = i915_mmio_reg_offset(HSW_CS_GPR(JUMP_PREDICATE));
+	*cs++ = i915_mmio_reg_offset(MI_PREDICATE_RESULT_1);
+
+	/* Restart from the beginning if we had timestamps roll over. */
+	*cs++ = (INTEL_GEN(i915) < 8 ?
+		 MI_BATCH_BUFFER_START : MI_BATCH_BUFFER_START_GEN8) |
+		MI_BATCH_PREDICATE;
+	*cs++ = i915_ggtt_offset(vma) + (ts0 - batch) * 4;
+	*cs++ = 0;
+
+	/*
+	 * Now add the diff between to previous timestamps and add it to :
+	 *      (((1 * << 64) - 1) - delay_ns)
+	 *
+	 * When the Carry Flag contains 1 this means the elapsed time is
+	 * longer than the expected delay, and we can exit the wait loop.
+	 */
+	*cs++ = MI_LOAD_REGISTER_IMM(2);
+	*cs++ = i915_mmio_reg_offset(HSW_CS_GPR(DELTA_TARGET));
+	*cs++ = lower_32_bits(delay_ticks);
+	*cs++ = i915_mmio_reg_offset(HSW_CS_GPR(DELTA_TARGET)) + 4;
+	*cs++ = upper_32_bits(delay_ticks);
+
+	*cs++ = MI_MATH(4);
+	*cs++ = MI_ALU_OP(MI_ALU_OP_LOAD, MI_ALU_SRC_SRCA, MI_ALU_SRC_REG(DELTA_TS));
+	*cs++ = MI_ALU_OP(MI_ALU_OP_LOAD, MI_ALU_SRC_SRCB, MI_ALU_SRC_REG(DELTA_TARGET));
+	*cs++ = MI_ALU_OP(MI_ALU_OP_ADD, 0, 0);
+	*cs++ = MI_ALU_OP(MI_ALU_OP_STOREINV, MI_ALU_SRC_REG(JUMP_PREDICATE), MI_ALU_SRC_CF);
+
+	/*
+	 * Transfer the result into the predicate register to be used for the
+	 * predicated jump.
+	 */
+	*cs++ = MI_LOAD_REGISTER_REG | (3 - 2);
+	*cs++ = i915_mmio_reg_offset(HSW_CS_GPR(JUMP_PREDICATE));
+	*cs++ = i915_mmio_reg_offset(MI_PREDICATE_RESULT_1);
+
+	/* Predicate the jump.  */
+	*cs++ = (INTEL_GEN(i915) < 8 ?
+		 MI_BATCH_BUFFER_START : MI_BATCH_BUFFER_START_GEN8) |
+		MI_BATCH_PREDICATE;
+	*cs++ = i915_ggtt_offset(vma) + (jump - batch) * 4;
+	*cs++ = 0;
+
+	/* Restore registers. */
+	for (i = 0; i < N_CS_GPR; i++) {
+		cs = save_restore_register(
+			i915, cs, false /* save */, HSW_CS_GPR(i),
+			INTEL_GT_SCRATCH_FIELD_PERF_CS_GPR + 8 * i, 2);
+	}
+	cs = save_restore_register(
+		i915, cs, false /* save */, MI_PREDICATE_RESULT_1,
+		INTEL_GT_SCRATCH_FIELD_PERF_PREDICATE_RESULT_1, 1);
+
+	/* And return to the ring. */
+	*cs++ = MI_BATCH_BUFFER_END;
+
+	GEM_BUG_ON((cs - batch) > (PAGE_SIZE / sizeof(*batch)));
+
+	i915_gem_object_flush_map(bo);
+	i915_gem_object_unpin_map(bo);
+
+	stream->noa_wait = vma;
+
+	return 0;
+
+err_unpin:
+	__i915_vma_unpin(vma);
+
+err_unref:
+	i915_gem_object_put(bo);
+
+	return ret;
+}
+
 static void config_oa_regs(struct drm_i915_private *dev_priv,
 			   const struct i915_oa_reg *regs,
 			   u32 n_regs)
@@ -2403,6 +2619,12 @@  static int i915_oa_stream_init(struct i915_perf_stream *stream,
 		}
 	}
 
+	ret = alloc_noa_wait(stream);
+	if (ret) {
+		DRM_DEBUG("Unable to allocate NOA wait batch buffer\n");
+		goto err_noa_wait_alloc;
+	}
+
 	ret = i915_perf_get_oa_config(dev_priv, props->metrics_set,
 				      &stream->oa_config);
 	if (ret) {
@@ -2469,6 +2691,9 @@  static int i915_oa_stream_init(struct i915_perf_stream *stream,
 	free_oa_configs(stream);
 
 err_config:
+	free_noa_wait(stream);
+
+err_noa_wait_alloc:
 	if (stream->ctx)
 		oa_put_render_ctx_id(stream);
 
@@ -3856,6 +4081,9 @@  void i915_perf_init(struct drm_i915_private *dev_priv)
 		ratelimit_set_flags(&dev_priv->perf.spurious_report_rs,
 				    RATELIMIT_MSG_ON_RELEASE);
 
+		atomic64_set(&dev_priv->perf.noa_programming_delay,
+			     500 * 1000 /* 500us */);
+
 		dev_priv->perf.initialized = true;
 	}
 }
diff --git a/drivers/gpu/drm/i915/i915_perf_types.h b/drivers/gpu/drm/i915/i915_perf_types.h
index b1c81107c062..695c6ad2fcd5 100644
--- a/drivers/gpu/drm/i915/i915_perf_types.h
+++ b/drivers/gpu/drm/i915/i915_perf_types.h
@@ -260,6 +260,12 @@  struct i915_perf_stream {
 		 */
 		u32 head;
 	} oa_buffer;
+
+	/**
+	 * A batch buffer doing a wait on the GPU for the NOA logic to be
+	 * reprogrammed.
+	 */
+	struct i915_vma *noa_wait;
 };
 
 /**
diff --git a/drivers/gpu/drm/i915/i915_reg.h b/drivers/gpu/drm/i915/i915_reg.h
index 006cffd56be2..8fea97b02365 100644
--- a/drivers/gpu/drm/i915/i915_reg.h
+++ b/drivers/gpu/drm/i915/i915_reg.h
@@ -545,7 +545,9 @@  static inline bool i915_mmio_reg_valid(i915_reg_t reg)
 #define MI_PREDICATE_SRC0_UDW	_MMIO(0x2400 + 4)
 #define MI_PREDICATE_SRC1	_MMIO(0x2408)
 #define MI_PREDICATE_SRC1_UDW	_MMIO(0x2408 + 4)
-
+#define MI_PREDICATE_DATA       _MMIO(0x2410)
+#define MI_PREDICATE_RESULT     _MMIO(0x2418)
+#define MI_PREDICATE_RESULT_1   _MMIO(0x241c)
 #define MI_PREDICATE_RESULT_2	_MMIO(0x2214)
 #define  LOWER_SLICE_ENABLED	(1 << 0)
 #define  LOWER_SLICE_DISABLED	(0 << 0)